git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_cache.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_cache.h"
   5 #include "btree_io.h"
   6 #include "btree_iter.h"
   7 #include "btree_locking.h"
   8 #include "debug.h"
   9
  10 #include <linux/prefetch.h>
  11 #include <linux/sched/mm.h>
  12 #include <trace/events/bcachefs.h>
  13
  14 const char * const bch2_btree_ids[] = {
  15 #define x(kwd, val, name) name,
  16         BCH_BTREE_IDS()
  17 #undef x
  18         NULL
  19 };
  20
  21 void bch2_recalc_btree_reserve(struct bch_fs *c)
  22 {
  23         unsigned i, reserve = 16;
  24
  25         if (!c->btree_roots[0].b)
  26                 reserve += 8;
  27
  28         for (i = 0; i < BTREE_ID_NR; i++)
  29                 if (c->btree_roots[i].b)
  30                         reserve += min_t(unsigned, 1,
  31                                          c->btree_roots[i].b->level) * 8;
  32
  33         c->btree_cache.reserve = reserve;
  34 }
  35
  36 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  37 {
  38         return max_t(int, 0, bc->used - bc->reserve);
  39 }
  40
  41 static void __btree_node_data_free(struct bch_fs *c, struct btree *b)
  42 {
  43         EBUG_ON(btree_node_write_in_flight(b));
  44
  45         kvpfree(b->data, btree_bytes(c));
  46         b->data = NULL;
  47         bch2_btree_keys_free(b);
  48 }
  49
  50 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  51 {
  52         struct btree_cache *bc = &c->btree_cache;
  53
  54         __btree_node_data_free(c, b);
  55         bc->used--;
  56         list_move(&b->list, &bc->freed);
  57 }
  58
  59 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  60                                    const void *obj)
  61 {
  62         const struct btree *b = obj;
  63         const u64 *v = arg->key;
  64
  65         return PTR_HASH(&b->key) == *v ? 0 : 1;
  66 }
  67
  68 static const struct rhashtable_params bch_btree_cache_params = {
  69         .head_offset    = offsetof(struct btree, hash),
  70         .key_offset     = offsetof(struct btree, key.v),
  71         .key_len        = sizeof(struct bch_extent_ptr),
  72         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  73 };
  74
  75 static void btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  76 {
  77         struct btree_cache *bc = &c->btree_cache;
  78
  79         b->data = kvpmalloc(btree_bytes(c), gfp);
  80         if (!b->data)
  81                 goto err;
  82
  83         if (bch2_btree_keys_alloc(b, btree_page_order(c), gfp))
  84                 goto err;
  85
  86         bc->used++;
  87         list_move(&b->list, &bc->freeable);
  88         return;
  89 err:
  90         kvpfree(b->data, btree_bytes(c));
  91         b->data = NULL;
  92         list_move(&b->list, &bc->freed);
  93 }
  94
  95 static struct btree *btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
  96 {
  97         struct btree *b = kzalloc(sizeof(struct btree), gfp);
  98         if (!b)
  99                 return NULL;
 100
 101         bkey_btree_ptr_init(&b->key);
 102         six_lock_init(&b->lock);
 103         INIT_LIST_HEAD(&b->list);
 104         INIT_LIST_HEAD(&b->write_blocked);
 105
 106         btree_node_data_alloc(c, b, gfp);
 107         return b->data ? b : NULL;
 108 }
 109
 110 /* Btree in memory cache - hash table */
 111
 112 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 113 {
 114         rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 115
 116         /* Cause future lookups for this node to fail: */
 117         PTR_HASH(&b->key) = 0;
 118 }
 119
 120 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 121 {
 122         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 123                                              bch_btree_cache_params);
 124 }
 125
 126 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 127                                 unsigned level, enum btree_id id)
 128 {
 129         int ret;
 130
 131         b->level        = level;
 132         b->btree_id     = id;
 133
 134         mutex_lock(&bc->lock);
 135         ret = __bch2_btree_node_hash_insert(bc, b);
 136         if (!ret)
 137                 list_add(&b->list, &bc->live);
 138         mutex_unlock(&bc->lock);
 139
 140         return ret;
 141 }
 142
 143 __flatten
 144 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 145                                      const struct bkey_i *k)
 146 {
 147         return rhashtable_lookup_fast(&bc->table, &PTR_HASH(k),
 148                                       bch_btree_cache_params);
 149 }
 150
 151 /*
 152  * this version is for btree nodes that have already been freed (we're not
 153  * reaping a real btree node)
 154  */
 155 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 156 {
 157         struct btree_cache *bc = &c->btree_cache;
 158         int ret = 0;
 159
 160         lockdep_assert_held(&bc->lock);
 161
 162         if (!six_trylock_intent(&b->lock))
 163                 return -ENOMEM;
 164
 165         if (!six_trylock_write(&b->lock))
 166                 goto out_unlock_intent;
 167
 168         if (btree_node_noevict(b))
 169                 goto out_unlock;
 170
 171         if (!btree_node_may_write(b))
 172                 goto out_unlock;
 173
 174         if (btree_node_dirty(b) &&
 175             test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
 176                 goto out_unlock;
 177
 178         if (btree_node_dirty(b) ||
 179             btree_node_write_in_flight(b) ||
 180             btree_node_read_in_flight(b)) {
 181                 if (!flush)
 182                         goto out_unlock;
 183
 184                 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 185                                TASK_UNINTERRUPTIBLE);
 186
 187                 /*
 188                  * Using the underscore version because we don't want to compact
 189                  * bsets after the write, since this node is about to be evicted
 190                  * - unless btree verify mode is enabled, since it runs out of
 191                  * the post write cleanup:
 192                  */
 193                 if (verify_btree_ondisk(c))
 194                         bch2_btree_node_write(c, b, SIX_LOCK_intent);
 195                 else
 196                         __bch2_btree_node_write(c, b, SIX_LOCK_read);
 197
 198                 /* wait for any in flight btree write */
 199                 btree_node_wait_on_io(b);
 200         }
 201 out:
 202         if (PTR_HASH(&b->key) && !ret)
 203                 trace_btree_node_reap(c, b);
 204         return ret;
 205 out_unlock:
 206         six_unlock_write(&b->lock);
 207 out_unlock_intent:
 208         six_unlock_intent(&b->lock);
 209         ret = -ENOMEM;
 210         goto out;
 211 }
 212
 213 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 214 {
 215         return __btree_node_reclaim(c, b, false);
 216 }
 217
 218 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 219 {
 220         return __btree_node_reclaim(c, b, true);
 221 }
 222
 223 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 224                                            struct shrink_control *sc)
 225 {
 226         struct bch_fs *c = container_of(shrink, struct bch_fs,
 227                                         btree_cache.shrink);
 228         struct btree_cache *bc = &c->btree_cache;
 229         struct btree *b, *t;
 230         unsigned long nr = sc->nr_to_scan;
 231         unsigned long can_free;
 232         unsigned long touched = 0;
 233         unsigned long freed = 0;
 234         unsigned i;
 235
 236         if (btree_shrinker_disabled(c))
 237                 return SHRINK_STOP;
 238
 239         /* Return -1 if we can't do anything right now */
 240         if (sc->gfp_mask & __GFP_IO)
 241                 mutex_lock(&bc->lock);
 242         else if (!mutex_trylock(&bc->lock))
 243                 return -1;
 244
 245         /*
 246          * It's _really_ critical that we don't free too many btree nodes - we
 247          * have to always leave ourselves a reserve. The reserve is how we
 248          * guarantee that allocating memory for a new btree node can always
 249          * succeed, so that inserting keys into the btree can always succeed and
 250          * IO can always make forward progress:
 251          */
 252         nr /= btree_pages(c);
 253         can_free = btree_cache_can_free(bc);
 254         nr = min_t(unsigned long, nr, can_free);
 255
 256         i = 0;
 257         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 258                 touched++;
 259
 260                 if (freed >= nr)
 261                         break;
 262
 263                 if (++i > 3 &&
 264                     !btree_node_reclaim(c, b)) {
 265                         btree_node_data_free(c, b);
 266                         six_unlock_write(&b->lock);
 267                         six_unlock_intent(&b->lock);
 268                         freed++;
 269                 }
 270         }
 271 restart:
 272         list_for_each_entry_safe(b, t, &bc->live, list) {
 273                 touched++;
 274
 275                 if (freed >= nr) {
 276                         /* Save position */
 277                         if (&t->list != &bc->live)
 278                                 list_move_tail(&bc->live, &t->list);
 279                         break;
 280                 }
 281
 282                 if (!btree_node_accessed(b) &&
 283                     !btree_node_reclaim(c, b)) {
 284                         /* can't call bch2_btree_node_hash_remove under lock  */
 285                         freed++;
 286                         if (&t->list != &bc->live)
 287                                 list_move_tail(&bc->live, &t->list);
 288
 289                         btree_node_data_free(c, b);
 290                         mutex_unlock(&bc->lock);
 291
 292                         bch2_btree_node_hash_remove(bc, b);
 293                         six_unlock_write(&b->lock);
 294                         six_unlock_intent(&b->lock);
 295
 296                         if (freed >= nr)
 297                                 goto out;
 298
 299                         if (sc->gfp_mask & __GFP_IO)
 300                                 mutex_lock(&bc->lock);
 301                         else if (!mutex_trylock(&bc->lock))
 302                                 goto out;
 303                         goto restart;
 304                 } else
 305                         clear_btree_node_accessed(b);
 306         }
 307
 308         mutex_unlock(&bc->lock);
 309 out:
 310         return (unsigned long) freed * btree_pages(c);
 311 }
 312
 313 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 314                                             struct shrink_control *sc)
 315 {
 316         struct bch_fs *c = container_of(shrink, struct bch_fs,
 317                                         btree_cache.shrink);
 318         struct btree_cache *bc = &c->btree_cache;
 319
 320         if (btree_shrinker_disabled(c))
 321                 return 0;
 322
 323         return btree_cache_can_free(bc) * btree_pages(c);
 324 }
 325
 326 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 327 {
 328         struct btree_cache *bc = &c->btree_cache;
 329         struct btree *b;
 330         unsigned i;
 331
 332         if (bc->shrink.list.next)
 333                 unregister_shrinker(&bc->shrink);
 334
 335         mutex_lock(&bc->lock);
 336
 337 #ifdef CONFIG_BCACHEFS_DEBUG
 338         if (c->verify_data)
 339                 list_move(&c->verify_data->list, &bc->live);
 340
 341         kvpfree(c->verify_ondisk, btree_bytes(c));
 342 #endif
 343
 344         for (i = 0; i < BTREE_ID_NR; i++)
 345                 if (c->btree_roots[i].b)
 346                         list_add(&c->btree_roots[i].b->list, &bc->live);
 347
 348         list_splice(&bc->freeable, &bc->live);
 349
 350         while (!list_empty(&bc->live)) {
 351                 b = list_first_entry(&bc->live, struct btree, list);
 352
 353                 BUG_ON(btree_node_read_in_flight(b) ||
 354                        btree_node_write_in_flight(b));
 355
 356                 if (btree_node_dirty(b))
 357                         bch2_btree_complete_write(c, b, btree_current_write(b));
 358                 clear_btree_node_dirty(b);
 359
 360                 btree_node_data_free(c, b);
 361         }
 362
 363         while (!list_empty(&bc->freed)) {
 364                 b = list_first_entry(&bc->freed, struct btree, list);
 365                 list_del(&b->list);
 366                 kfree(b);
 367         }
 368
 369         mutex_unlock(&bc->lock);
 370
 371         if (bc->table_init_done)
 372                 rhashtable_destroy(&bc->table);
 373 }
 374
 375 int bch2_fs_btree_cache_init(struct bch_fs *c)
 376 {
 377         struct btree_cache *bc = &c->btree_cache;
 378         unsigned i;
 379         int ret = 0;
 380
 381         pr_verbose_init(c->opts, "");
 382
 383         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 384         if (ret)
 385                 goto out;
 386
 387         bc->table_init_done = true;
 388
 389         bch2_recalc_btree_reserve(c);
 390
 391         for (i = 0; i < bc->reserve; i++)
 392                 if (!btree_node_mem_alloc(c, GFP_KERNEL)) {
 393                         ret = -ENOMEM;
 394                         goto out;
 395                 }
 396
 397         list_splice_init(&bc->live, &bc->freeable);
 398
 399 #ifdef CONFIG_BCACHEFS_DEBUG
 400         mutex_init(&c->verify_lock);
 401
 402         c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
 403         if (!c->verify_ondisk) {
 404                 ret = -ENOMEM;
 405                 goto out;
 406         }
 407
 408         c->verify_data = btree_node_mem_alloc(c, GFP_KERNEL);
 409         if (!c->verify_data) {
 410                 ret = -ENOMEM;
 411                 goto out;
 412         }
 413
 414         list_del_init(&c->verify_data->list);
 415 #endif
 416
 417         bc->shrink.count_objects        = bch2_btree_cache_count;
 418         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 419         bc->shrink.seeks                = 4;
 420         bc->shrink.batch                = btree_pages(c) * 2;
 421         register_shrinker(&bc->shrink);
 422 out:
 423         pr_verbose_init(c->opts, "ret %i", ret);
 424         return ret;
 425 }
 426
 427 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 428 {
 429         mutex_init(&bc->lock);
 430         INIT_LIST_HEAD(&bc->live);
 431         INIT_LIST_HEAD(&bc->freeable);
 432         INIT_LIST_HEAD(&bc->freed);
 433 }
 434
 435 /*
 436  * We can only have one thread cannibalizing other cached btree nodes at a time,
 437  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 438  * cannibalize_bucket() will take. This means every time we unlock the root of
 439  * the btree, we need to release this lock if we have it held.
 440  */
 441 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 442 {
 443         struct btree_cache *bc = &c->btree_cache;
 444
 445         if (bc->alloc_lock == current) {
 446                 trace_btree_node_cannibalize_unlock(c);
 447                 bc->alloc_lock = NULL;
 448                 closure_wake_up(&bc->alloc_wait);
 449         }
 450 }
 451
 452 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 453 {
 454         struct btree_cache *bc = &c->btree_cache;
 455         struct task_struct *old;
 456
 457         old = cmpxchg(&bc->alloc_lock, NULL, current);
 458         if (old == NULL || old == current)
 459                 goto success;
 460
 461         if (!cl) {
 462                 trace_btree_node_cannibalize_lock_fail(c);
 463                 return -ENOMEM;
 464         }
 465
 466         closure_wait(&bc->alloc_wait, cl);
 467
 468         /* Try again, after adding ourselves to waitlist */
 469         old = cmpxchg(&bc->alloc_lock, NULL, current);
 470         if (old == NULL || old == current) {
 471                 /* We raced */
 472                 closure_wake_up(&bc->alloc_wait);
 473                 goto success;
 474         }
 475
 476         trace_btree_node_cannibalize_lock_fail(c);
 477         return -EAGAIN;
 478
 479 success:
 480         trace_btree_node_cannibalize_lock(c);
 481         return 0;
 482 }
 483
 484 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 485 {
 486         struct btree_cache *bc = &c->btree_cache;
 487         struct btree *b;
 488
 489         list_for_each_entry_reverse(b, &bc->live, list)
 490                 if (!btree_node_reclaim(c, b))
 491                         return b;
 492
 493         while (1) {
 494                 list_for_each_entry_reverse(b, &bc->live, list)
 495                         if (!btree_node_write_and_reclaim(c, b))
 496                                 return b;
 497
 498                 /*
 499                  * Rare case: all nodes were intent-locked.
 500                  * Just busy-wait.
 501                  */
 502                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 503                 cond_resched();
 504         }
 505 }
 506
 507 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 508 {
 509         struct btree_cache *bc = &c->btree_cache;
 510         struct btree *b;
 511         u64 start_time = local_clock();
 512         unsigned flags;
 513
 514         flags = memalloc_nofs_save();
 515         mutex_lock(&bc->lock);
 516
 517         /*
 518          * btree_free() doesn't free memory; it sticks the node on the end of
 519          * the list. Check if there's any freed nodes there:
 520          */
 521         list_for_each_entry(b, &bc->freeable, list)
 522                 if (!btree_node_reclaim(c, b))
 523                         goto out_unlock;
 524
 525         /*
 526          * We never free struct btree itself, just the memory that holds the on
 527          * disk node. Check the freed list before allocating a new one:
 528          */
 529         list_for_each_entry(b, &bc->freed, list)
 530                 if (!btree_node_reclaim(c, b)) {
 531                         btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_NOIO);
 532                         if (b->data)
 533                                 goto out_unlock;
 534
 535                         six_unlock_write(&b->lock);
 536                         six_unlock_intent(&b->lock);
 537                         goto err;
 538                 }
 539
 540         b = btree_node_mem_alloc(c, __GFP_NOWARN|GFP_NOIO);
 541         if (!b)
 542                 goto err;
 543
 544         BUG_ON(!six_trylock_intent(&b->lock));
 545         BUG_ON(!six_trylock_write(&b->lock));
 546 out_unlock:
 547         BUG_ON(btree_node_hashed(b));
 548         BUG_ON(btree_node_write_in_flight(b));
 549
 550         list_del_init(&b->list);
 551         mutex_unlock(&bc->lock);
 552         memalloc_nofs_restore(flags);
 553 out:
 554         b->flags                = 0;
 555         b->written              = 0;
 556         b->nsets                = 0;
 557         b->sib_u64s[0]          = 0;
 558         b->sib_u64s[1]          = 0;
 559         b->whiteout_u64s        = 0;
 560         b->uncompacted_whiteout_u64s = 0;
 561         bch2_btree_keys_init(b, &c->expensive_debug_checks);
 562
 563         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 564                                start_time);
 565
 566         return b;
 567 err:
 568         /* Try to cannibalize another cached btree node: */
 569         if (bc->alloc_lock == current) {
 570                 b = btree_node_cannibalize(c);
 571                 list_del_init(&b->list);
 572                 mutex_unlock(&bc->lock);
 573
 574                 bch2_btree_node_hash_remove(bc, b);
 575
 576                 trace_btree_node_cannibalize(c);
 577                 goto out;
 578         }
 579
 580         mutex_unlock(&bc->lock);
 581         return ERR_PTR(-ENOMEM);
 582 }
 583
 584 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 585 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 586                                 struct btree_iter *iter,
 587                                 const struct bkey_i *k,
 588                                 unsigned level,
 589                                 enum six_lock_type lock_type,
 590                                 bool sync)
 591 {
 592         struct btree_cache *bc = &c->btree_cache;
 593         struct btree *b;
 594
 595         /*
 596          * Parent node must be locked, else we could read in a btree node that's
 597          * been freed:
 598          */
 599         BUG_ON(!btree_node_locked(iter, level + 1));
 600         BUG_ON(level >= BTREE_MAX_DEPTH);
 601
 602         b = bch2_btree_node_mem_alloc(c);
 603         if (IS_ERR(b))
 604                 return b;
 605
 606         bkey_copy(&b->key, k);
 607         if (bch2_btree_node_hash_insert(bc, b, level, iter->btree_id)) {
 608                 /* raced with another fill: */
 609
 610                 /* mark as unhashed... */
 611                 PTR_HASH(&b->key) = 0;
 612
 613                 mutex_lock(&bc->lock);
 614                 list_add(&b->list, &bc->freeable);
 615                 mutex_unlock(&bc->lock);
 616
 617                 six_unlock_write(&b->lock);
 618                 six_unlock_intent(&b->lock);
 619                 return NULL;
 620         }
 621
 622         /*
 623          * If the btree node wasn't cached, we can't drop our lock on
 624          * the parent until after it's added to the cache - because
 625          * otherwise we could race with a btree_split() freeing the node
 626          * we're trying to lock.
 627          *
 628          * But the deadlock described below doesn't exist in this case,
 629          * so it's safe to not drop the parent lock until here:
 630          */
 631         if (btree_node_read_locked(iter, level + 1))
 632                 btree_node_unlock(iter, level + 1);
 633
 634         bch2_btree_node_read(c, b, sync);
 635
 636         six_unlock_write(&b->lock);
 637
 638         if (!sync) {
 639                 six_unlock_intent(&b->lock);
 640                 return NULL;
 641         }
 642
 643         if (lock_type == SIX_LOCK_read)
 644                 six_lock_downgrade(&b->lock);
 645
 646         return b;
 647 }
 648
 649 /**
 650  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 651  * in from disk if necessary.
 652  *
 653  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 654  *
 655  * The btree node will have either a read or a write lock held, depending on
 656  * the @write parameter.
 657  */
 658 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 659                                   const struct bkey_i *k, unsigned level,
 660                                   enum six_lock_type lock_type)
 661 {
 662         struct btree_cache *bc = &c->btree_cache;
 663         struct btree *b;
 664         struct bset_tree *t;
 665
 666         /*
 667          * XXX: locking optimization
 668          *
 669          * we can make the locking looser here - caller can drop lock on parent
 670          * node before locking child node (and potentially blocking): we just
 671          * have to have bch2_btree_node_fill() call relock on the parent and
 672          * return -EINTR if that fails
 673          */
 674         EBUG_ON(!btree_node_locked(iter, level + 1));
 675         EBUG_ON(level >= BTREE_MAX_DEPTH);
 676 retry:
 677         b = btree_cache_find(bc, k);
 678         if (unlikely(!b)) {
 679                 /*
 680                  * We must have the parent locked to call bch2_btree_node_fill(),
 681                  * else we could read in a btree node from disk that's been
 682                  * freed:
 683                  */
 684                 b = bch2_btree_node_fill(c, iter, k, level, lock_type, true);
 685
 686                 /* We raced and found the btree node in the cache */
 687                 if (!b)
 688                         goto retry;
 689
 690                 if (IS_ERR(b))
 691                         return b;
 692         } else {
 693                 /*
 694                  * There's a potential deadlock with splits and insertions into
 695                  * interior nodes we have to avoid:
 696                  *
 697                  * The other thread might be holding an intent lock on the node
 698                  * we want, and they want to update its parent node so they're
 699                  * going to upgrade their intent lock on the parent node to a
 700                  * write lock.
 701                  *
 702                  * But if we're holding a read lock on the parent, and we're
 703                  * trying to get the intent lock they're holding, we deadlock.
 704                  *
 705                  * So to avoid this we drop the read locks on parent nodes when
 706                  * we're starting to take intent locks - and handle the race.
 707                  *
 708                  * The race is that they might be about to free the node we
 709                  * want, and dropping our read lock on the parent node lets them
 710                  * update the parent marking the node we want as freed, and then
 711                  * free it:
 712                  *
 713                  * To guard against this, btree nodes are evicted from the cache
 714                  * when they're freed - and PTR_HASH() is zeroed out, which we
 715                  * check for after we lock the node.
 716                  *
 717                  * Then, bch2_btree_node_relock() on the parent will fail - because
 718                  * the parent was modified, when the pointer to the node we want
 719                  * was removed - and we'll bail out:
 720                  */
 721                 if (btree_node_read_locked(iter, level + 1))
 722                         btree_node_unlock(iter, level + 1);
 723
 724                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type))
 725                         return ERR_PTR(-EINTR);
 726
 727                 if (unlikely(PTR_HASH(&b->key) != PTR_HASH(k) ||
 728                              b->level != level ||
 729                              race_fault())) {
 730                         six_unlock_type(&b->lock, lock_type);
 731                         if (bch2_btree_node_relock(iter, level + 1))
 732                                 goto retry;
 733
 734                         trace_trans_restart_btree_node_reused(iter->trans->ip);
 735                         return ERR_PTR(-EINTR);
 736                 }
 737         }
 738
 739         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 740                        TASK_UNINTERRUPTIBLE);
 741
 742         prefetch(b->aux_data);
 743
 744         for_each_bset(b, t) {
 745                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 746
 747                 prefetch(p + L1_CACHE_BYTES * 0);
 748                 prefetch(p + L1_CACHE_BYTES * 1);
 749                 prefetch(p + L1_CACHE_BYTES * 2);
 750         }
 751
 752         /* avoid atomic set bit if it's not needed: */
 753         if (btree_node_accessed(b))
 754                 set_btree_node_accessed(b);
 755
 756         if (unlikely(btree_node_read_error(b))) {
 757                 six_unlock_type(&b->lock, lock_type);
 758                 return ERR_PTR(-EIO);
 759         }
 760
 761         EBUG_ON(b->btree_id != iter->btree_id ||
 762                 BTREE_NODE_LEVEL(b->data) != level ||
 763                 bkey_cmp(b->data->max_key, k->k.p));
 764
 765         return b;
 766 }
 767
 768 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
 769                                           struct btree_iter *iter,
 770                                           struct btree *b,
 771                                           enum btree_node_sibling sib)
 772 {
 773         struct btree_trans *trans = iter->trans;
 774         struct btree *parent;
 775         struct btree_node_iter node_iter;
 776         struct bkey_packed *k;
 777         BKEY_PADDED(k) tmp;
 778         struct btree *ret = NULL;
 779         unsigned level = b->level;
 780
 781         parent = btree_iter_node(iter, level + 1);
 782         if (!parent)
 783                 return NULL;
 784
 785         if (!bch2_btree_node_relock(iter, level + 1)) {
 786                 ret = ERR_PTR(-EINTR);
 787                 goto out;
 788         }
 789
 790         node_iter = iter->l[parent->level].iter;
 791
 792         k = bch2_btree_node_iter_peek_all(&node_iter, parent);
 793         BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
 794
 795         k = sib == btree_prev_sib
 796                 ? bch2_btree_node_iter_prev(&node_iter, parent)
 797                 : (bch2_btree_node_iter_advance(&node_iter, parent),
 798                    bch2_btree_node_iter_peek(&node_iter, parent));
 799         if (!k)
 800                 goto out;
 801
 802         bch2_bkey_unpack(parent, &tmp.k, k);
 803
 804         ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 805                                   SIX_LOCK_intent);
 806
 807         if (PTR_ERR_OR_ZERO(ret) == -EINTR && !trans->nounlock) {
 808                 struct btree_iter *linked;
 809
 810                 if (!bch2_btree_node_relock(iter, level + 1))
 811                         goto out;
 812
 813                 /*
 814                  * We might have got -EINTR because trylock failed, and we're
 815                  * holding other locks that would cause us to deadlock:
 816                  */
 817                 trans_for_each_iter(trans, linked)
 818                         if (btree_iter_cmp(iter, linked) < 0)
 819                                 __bch2_btree_iter_unlock(linked);
 820
 821                 if (sib == btree_prev_sib)
 822                         btree_node_unlock(iter, level);
 823
 824                 ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 825                                           SIX_LOCK_intent);
 826
 827                 /*
 828                  * before btree_iter_relock() calls btree_iter_verify_locks():
 829                  */
 830                 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 831                         btree_node_unlock(iter, level + 1);
 832
 833                 if (!bch2_btree_node_relock(iter, level)) {
 834                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
 835
 836                         if (!IS_ERR(ret)) {
 837                                 six_unlock_intent(&ret->lock);
 838                                 ret = ERR_PTR(-EINTR);
 839                         }
 840                 }
 841
 842                 bch2_trans_relock(trans);
 843         }
 844 out:
 845         if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 846                 btree_node_unlock(iter, level + 1);
 847
 848         if (PTR_ERR_OR_ZERO(ret) == -EINTR)
 849                 bch2_btree_iter_upgrade(iter, level + 2);
 850
 851         BUG_ON(!IS_ERR(ret) && !btree_node_locked(iter, level));
 852
 853         if (!IS_ERR_OR_NULL(ret)) {
 854                 struct btree *n1 = ret, *n2 = b;
 855
 856                 if (sib != btree_prev_sib)
 857                         swap(n1, n2);
 858
 859                 BUG_ON(bkey_cmp(btree_type_successor(n1->btree_id,
 860                                                      n1->key.k.p),
 861                                 n2->data->min_key));
 862         }
 863
 864         bch2_btree_trans_verify_locks(trans);
 865
 866         return ret;
 867 }
 868
 869 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
 870                               const struct bkey_i *k, unsigned level)
 871 {
 872         struct btree_cache *bc = &c->btree_cache;
 873         struct btree *b;
 874
 875         BUG_ON(!btree_node_locked(iter, level + 1));
 876         BUG_ON(level >= BTREE_MAX_DEPTH);
 877
 878         b = btree_cache_find(bc, k);
 879         if (b)
 880                 return;
 881
 882         bch2_btree_node_fill(c, iter, k, level, SIX_LOCK_read, false);
 883 }
 884
 885 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
 886                              struct btree *b)
 887 {
 888         const struct bkey_format *f = &b->format;
 889         struct bset_stats stats;
 890
 891         memset(&stats, 0, sizeof(stats));
 892
 893         bch2_btree_keys_stats(b, &stats);
 894
 895         pr_buf(out,
 896                "l %u %llu:%llu - %llu:%llu:\n"
 897                "    ptrs: ",
 898                b->level,
 899                b->data->min_key.inode,
 900                b->data->min_key.offset,
 901                b->data->max_key.inode,
 902                b->data->max_key.offset);
 903         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 904         pr_buf(out, "\n"
 905                "    format: u64s %u fields %u %u %u %u %u\n"
 906                "    unpack fn len: %u\n"
 907                "    bytes used %zu/%zu (%zu%% full)\n"
 908                "    sib u64s: %u, %u (merge threshold %zu)\n"
 909                "    nr packed keys %u\n"
 910                "    nr unpacked keys %u\n"
 911                "    floats %zu\n"
 912                "    failed unpacked %zu\n"
 913                "    failed prev %zu\n"
 914                "    failed overflow %zu\n",
 915                f->key_u64s,
 916                f->bits_per_field[0],
 917                f->bits_per_field[1],
 918                f->bits_per_field[2],
 919                f->bits_per_field[3],
 920                f->bits_per_field[4],
 921                b->unpack_fn_len,
 922                b->nr.live_u64s * sizeof(u64),
 923                btree_bytes(c) - sizeof(struct btree_node),
 924                b->nr.live_u64s * 100 / btree_max_u64s(c),
 925                b->sib_u64s[0],
 926                b->sib_u64s[1],
 927                BTREE_FOREGROUND_MERGE_THRESHOLD(c),
 928                b->nr.packed_keys,
 929                b->nr.unpacked_keys,
 930                stats.floats,
 931                stats.failed_unpacked,
 932                stats.failed_prev,
 933                stats.failed_overflow);
 934 }